Skip to content

0040-gpt-5-6-sol-on-design-planner-tiers.md

Latest commit

 

History

History
179 lines (144 loc) · 9.03 KB

0040-gpt-5-6-sol-on-design-planner-tiers.md

File metadata and controls

179 lines (144 loc) · 9.03 KB
Raw

0040. GPT-5.6 Sol on DESIGN+PLANNER Tiers and Implemented-by Commit Footer

opencode-era record. Superseded where moot by the pi migration (ADR-0055). Retained as historical context.

Date: 2026-07-27

Status

Accepted

References ADR-0031 (prior rebalance + footer rename — the precedent this extends), ADR-0030 (DESIGN tier), ADR-0013 (variant env var), ADR-0010 (footer/issue-ref convention). Does not supersede any ADR.

Context

The operator acquired a ChatGPT Plus subscription, giving access to GPT-5.6 Sol (OpenAI's flagship reasoning/coding model) through opencode's first-class ChatGPT-Plus/Pro OAuth provider — a path OpenAI explicitly permits for third-party developer tooling (unlike Anthropic, which prohibits the equivalent). opencode authenticates via /connect → OpenAI → "ChatGPT Plus/Pro"; no API key, no per-token billing. The binding economic constraint is therefore the ChatGPT Plus weekly window (surfaced by the @slkiser/opencode-quota plugin), not pay-per-token cost.

This is the first harness tier backed by a subscription-OAuth auth model rather than an API key. It also creates a footer-attribution gap: ADR-0031 §4 sourced Authored-by: from agent.plan.model (PLANNER) on the rationale that the creation pipeline (design → plan → build) was "all GLM". Once PLANNER moves to Sol while PRIMARY (build/tdd) stays on GLM, that rationale breaks — a @tdd-authored commit would carry Authored-by: gpt-5.6-sol, mislabelling GLM-authored code. The @architect review flagged this as its leading condition.

Two governing principles:

  1. Maximize quality-per-quota on high-leverage thinking — DESIGN/PLANNER are low/medium-frequency, high-leverage tiers; Sol earns its place there.
  2. Attribute every model in the pipeline — add an Implemented-by: footer sourced from PRIMARY so Sol (plan), GLM (code), and DeepSeek (review) are each represented, resolving the ADR-0031 §4 drift.

Decision

1. Route GPT-5.6 Sol to DESIGN + PLANNER at xhigh

Tier Model Variant Agents
PRIMARY zai-coding-plan/glm-5.2 max build, tdd, debug, resolve-merge-conflicts, general
PLANNER openai/gpt-5.6-sol xhigh plan, from-issue, architect, consult
DESIGN openai/gpt-5.6-sol xhigh design
JUDGE deepseek/deepseek-v4-pro medium code-review, standards-review, spec-review, test-audit, judge, explore
UTILITY deepseek/deepseek-v4-flash medium compaction, title, summary, docs-writer, semgrep

Only setup.json values change. The {env:VAR} indirection (ADR-0012/0022) means no opencode.jsonc, .envrc, or agent-tier-membership edit. No agent moves tier (unlike ADR-0031's 7-agent reassignment).

2. ChatGPT-Plus OAuth auth path; weekly-window economics

Tier placement is decided by call frequency × downstream leverage, not per-token price. PRIMARY (highest frequency: every build turn, every @tdd iteration) is excluded as the burn risk; Sol stays reachable ad-hoc via /models. DESIGN (once per brainstorm) and PLANNER (once per feature) are the defensible fit — high leverage, low window consumption. PRIMARY stays on high-capacity flat-rate GLM (~166M tokens/week, ~41% utilized per ADR-0031), which never hits the wall.

3. xhigh over a custom max variant

opencode's built-in OpenAI variants top at xhigh (none · minimal · low · medium · high · xhigh); the API exposes a higher max that opencode does not ship built-in. We choose the built-in xhigh for zero-config reliability over the absolute peak. Per ADR-0031's "variant values are provider-relative" principle, this is not a downgrade of intent: both GLM max (ExtraHigh) and OpenAI xhigh are their provider's opencode-native reasoning ceiling among shipped variants. The absolute-ceiling path (custom max variant) is the documented fallback (spec §2.1 option B).

4. Cross-model review preserved and enriched

JUDGE stays DeepSeek. Post-change the generators span two providers (OpenAI on DESIGN/PLANNER, GLM on PRIMARY) and the reviewer is a third (DeepSeek), so DeepSeek now cross-reviews both OpenAI- and GLM-authored work. Moving JUDGE to Sol would collapse the cross-model property into a self-review.

5. Operating model: manual /models fallback

When the weekly window runs low, the operator switches DESIGN/PLANNER sessions back to GLM via /models for the remainder of the window. The @slkiser/opencode-quota toast/sidebar surfaces the window. No auto-fallback is engineered (YAGNI — opencode exposes no hook for it; two low-frequency tiers don't justify one).

6. Implemented-by: commit footer (resolves the ADR-0031 §4 drift)

Add a fourth required trailer sourced from the PRIMARY tier:

Footer Tier source Resolves to Stage
Authored-by: PLANNER (agent.plan.model) gpt-5.6-sol design/plan
Implemented-by: PRIMARY (agent.tdd.model / agent.build.model inherit {env:OPENCODE_MODEL_PRIMARY}) glm-5.2 code
Tested-by: JUDGE (agent.code-review.model) deepseek-v4-pro review
Signed-off-by: user (resolve-identity.sh) kyau <git@kyaulabs.com> human

Pipeline order: Refs:/Fixes: → Authored-by: → Implemented-by: → Tested-by: → Signed-off-by:. The commitlint trailers-exist rule is extended to require Implemented-by:; the issue-ref-convention rule (Refs/Fixes before Authored-by) is unchanged. Like the existing footers, Implemented-by: is a fixed-source trailer applied uniformly regardless of which agent authors the commit (dynamic per-author re-sourcing is explicitly out of scope). This extends ADR-0031's footer-rename lineage and leaves ADR-0010's ordering rule unaffected.

Consequences

Positive:

  • The strongest available model drives the two tiers where a single turn shapes the most downstream work (specs, plans).
  • All three models in the pipeline are attributed in commit metadata (Sol/GLM/DeepSeek) — one footer per stage.
  • The ADR-0031 §4 Authored-by: semantic drift is resolved (not merely documented) by the Implemented-by: addition.
  • Cross-model review is enriched (DeepSeek reviews two generator providers).
  • The change is mechanically minimal: 2 setup.json values flow through existing {env:VAR} indirection; no structural edit.

Negative:

  • The Plus weekly window is a qualitative budget (not a published token count); exhaustion during heavy DESIGN/PLANNER use requires manual /models fallback.
  • Implemented-by: (like the existing footers) is mechanically sourced from PRIMARY regardless of authoring agent, so a design-agent docs commit carries Implemented-by: glm-5.2 even though no code was implemented — the same fixed-source imprecision the harness already accepts for Authored-by/ Tested-by on docs commits.
  • The commit-msg_test.sh footer-fixture ripple (13 fixtures) adds test maintenance surface, though breakage is detection-easy (commitlint rejects 3-trailer messages noisily).

Neutral:

  • PRIMARY, JUDGE, and UTILITY tiers are unchanged.
  • The OpenAI API-key billing path is unused (irrelevant under ChatGPT-Plus auth).
  • ADR-0031 is referenced, not superseded.

Alternatives Considered

  • Put Sol on PRIMARY (the highest-volume tier): Rejected. PRIMARY is the burn risk on a weekly window; defaulting it to Sol would exhaust the window during heavy @tdd sprints. Sol remains reachable ad-hoc via /models.
  • Put Sol on all thinking + coding tiers (DESIGN+PLANNER+PRIMARY): Rejected. Same burn-risk problem, amplified.
  • Move JUDGE to Sol: Rejected. Collapses the cross-model review property into a self-review.
  • Custom max variant over built-in xhigh: Rejected as the default. Preserves the literal max string and hits Sol's absolute ceiling, but adds a config block and depends on the AI SDK passing reasoningEffort:"max" through (unverified). Kept as the documented fallback.
  • Document the Authored-by: drift as a known limitation (the architect's fallback): Rejected in favour of the active fix. Adding Implemented-by: resolves the drift; mere documentation would leave commit metadata inaccurate.
  • Dynamic per-commit footer re-sourcing (footer = actual authoring agent's model): Rejected. A different model from the harness's fixed-source footer convention; retained fixed-source for consistency.

Cross-references

  • ADR-0031 (referenced — prior rebalance + footer rename; the precedent extended)
  • ADR-0030 (referenced — DESIGN tier independent-configurability rationale, exercised here)
  • ADR-0013 (not contradicted — variant env-var mechanism preserved)
  • ADR-0012 (not contradicted — {env:VAR} pattern preserved)
  • ADR-0022 (not contradicted — agent config in opencode.jsonc preserved)
  • ADR-0029 (not contradicted — setup.json structure preserved)
  • ADR-0010 (not contradicted — issue-ref-before-Authored-by ordering preserved)
  • Spec: docs/specs/2026-07-27-gpt-5-6-sol-design-planner-tiers-spec.md
  • Plan: docs/plans/2026-07-27-gpt-5-6-sol-design-planner-tiers.md